AI Safety & Governance · Projected milestone
Weak AI checks strong AI
45% by 2033 · including earlier years
The year is an editorial timing estimate. The percentage applies to occurrence by the stated deadline. These estimates are not calibrated odds.
Scalable oversight methods let people and trusted weaker models reliably catch errors and deception in smarter models.
Explore this milestone on the interactive mapWhat the forecast means
Milestone stage: demonstration
partially-demonstrated
Weak-model monitoring works for some errors; robust adversarial oversight of stronger models is unconfirmed.
Completion criteria: Independent adversarial evaluation reports detection and false-positive rates, including attempts designed by the stronger model; success requires a pre-registered 95% recall on the stated consequential-error set at a usable review cost. Do not extrapolate to every possible behavior.
This page describes a forecast, rather than an achieved result. Read the methodology and assumptions.
Built on
Led to
Sources & evidence
- Measurements for understanding the pace of AI development inside frontier labs 2026-09-17
Anthropic reports AI leading 26% of R&D in August, no fully autonomous measured R&D subset, and about 30,000 concurrent internal agents. It reports complete action-monitor coverage, which is not complete detection. The measurements use internal/model judgments.
- Investigating unintended model actions 2026-10-09
Anthropic reports unintended actions in evaluation/internal settings and remedial controls. This is evidence of a control/reliability gap, not a measurement of the absence of underlying cognitive capability.
Record reviewed 2026-10-18